News

seguranca

CONFIRMED: Anthropic tightens defenses against copying Claude

Anthropic reports a 90% to 99% drop in data extracted from Claude. Its defenses also change integrations and charge for blocked requests, exposing tensions between safety, openness and control.

Ilustração editorial de um núcleo neural laranja protegido por um escudo transparente contra fluxos de dados, simbolizando as defesas do Claude contra destilação não autorizada.

Anthropic has tightened Claude's defenses against unauthorized copying campaigns and says it reduced the data extracted by 90% to 99%. Its October 12 announcement puts a tension at the center of the dispute: a company that publishes research and advocates external evaluation is also restricting how its model can be used. The protections affect legitimate customers, not just competitors accused of abuse.

The target is reasoning, not every form of distillation

Distilling a model means training another system on its responses. The technique is not inherently an attack. Anthropic is targeting industrial campaigns conducted without authorization, often supported by networks of fraudulent accounts. Claude's reasoning transcripts are particularly valuable because they teach another model how to reach an answer, rather than simply which answer to produce.

A dispute beyond intellectual property

The stated rationale combines safety and geopolitics. Anthropic says advanced capabilities can be copied without equivalent safeguards, increasing biological and cybersecurity risks, and incorporated into foreign military, intelligence and surveillance systems. In September, it attributed roughly 200 million interactions to five campaigns, including more than 151 million associated with Alibaba's Qwen, alongside operations attributed to Moonshot and DeepSeek.

In the case of Zhipu, also known as Z.ai, Anthropic accuses its rival of trying to extract Claude Fable 5 capabilities while developing GLM-5.3, from the General Language Model family. After encountering safeguards, Zhipu allegedly switched to an earlier Claude and another laboratory's model. The US Center for Advancing Innovation and Standards for Super Intelligence, or CAISSI, rated GLM-5.3 the most capable open weights model for cybersecurity released at that point. The claim that its safeguards are easily bypassed comes from Anthropic itself.

Another episode involves requests to Moonshot's Kimi that were allegedly secretly redirected to Claude. A user Anthropic believes was affiliated with China's People's Liberation Army requested analysis of closed circuit television footage, or CCTV, of a specific person. These attributions represent Anthropic's account of events, not independent findings established by this article.

What changes for Claude integrations

The first barrier changes the application programming interface, or API, used to connect Claude to other systems. With Claude Opus 5.5 and Claude Fable 5.1, accounts created on or after August 31, 2026 cannot edit prior context generated by Claude. Integrations that manipulated that history need to be reviewed.

The second layer is a two stage classifier that detects extraction, with a false positive rate below 0.1%, according to the company. The third imposes tighter usage limits and bans persistent offenders. Since September 24, requests blocked before a response have again incurred charges in covered areas, including distillation. Developers therefore need to monitor blocks and avoid repeated automatic attempts without investigating their cause.

The reported reduction has limits

The company measures extracted reasoning tokens, units of text processed by the model, on its own service. It reports a reduction by a factor of 10 to 100 and tens of thousands of banned accounts, often within two hours. It acknowledges lost revenue and trillions of tokens of rejected paying traffic, but does not know how long the gains will last. External coverage confirms the announcement; it does not independently audit these figures.

There is also a transparency tension. A Reuters investigation cited by The Deep Dive describes negotiations by Anthropic and OpenAI over restrictions on publishing external evaluations. Anthropic reserved report approval rights for some new evaluators, while saying established partners retain control over their findings. This does not invalidate the defenses, but it matters when assessing the independence of scrutiny.

The MaxAssistant view

Protecting capabilities may be necessary; demonstrating effectiveness requires more than a supplier's account. For Claude users, the immediate change is operational: review context handling, monitor blocks and include rejected requests in costs. Safety and commercial control coexist here, and neither removes the need for transparency.

Sources

Anthropic: Strengthening our defenses against distillation attacks; The Deep Dive: Anthropic Sacrifices Revenue to Protect Claude, While Keeping Approval Rights Over Some AI Audits; Blockchain.News: Anthropic Reports 90% to 99% Reduction in Distillation Attacks on Claude Models; IJR: Anthropic says safeguards cut Claude distillation 90% to 99%